Faster Deep & Rush

The first token now arrives 87% faster and entire responses are 32% faster, p50, in Amp's deep and rush modes. How? Mostly using websockets for communication with OpenAI, partly because we rebuilt Amp to be much faster last month. These gains matter most on long-horizon tasks, where we're seeing up to a 40% end-to-end speedup from user prompt submission to completion. You can see the difference: